projects / mesa.git / blob
? search:


a6981c136e70c5ce4ed2b832cb52804378f0f659
[mesa.git] / src / amd / vulkan / si_cmd_buffer.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based on si_state.c
6 * Copyright © 2015 Advanced Micro Devices, Inc.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 /* command buffer handling for SI */
29
30 #include "radv_private.h"
31 #include "radv_shader.h"
32 #include "radv_cs.h"
33 #include "sid.h"
34 #include "gfx9d.h"
35 #include "radv_util.h"
36 #include "main/macros.h"
37
38 static void
39 si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
40 struct radeon_winsys_cs *cs,
41 unsigned raster_config,
42 unsigned raster_config_1)
43 {
44 unsigned sh_per_se = MAX2(physical_device->rad_info.max_sh_per_se, 1);
45 unsigned num_se = MAX2(physical_device->rad_info.max_se, 1);
46 unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
47 unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
48 unsigned rb_per_pkr = MIN2(num_rb / num_se / sh_per_se, 2);
49 unsigned rb_per_se = num_rb / num_se;
50 unsigned se_mask[4];
51 unsigned se;
52
53 se_mask[0] = ((1 << rb_per_se) - 1) & rb_mask;
54 se_mask[1] = (se_mask[0] << rb_per_se) & rb_mask;
55 se_mask[2] = (se_mask[1] << rb_per_se) & rb_mask;
56 se_mask[3] = (se_mask[2] << rb_per_se) & rb_mask;
57
58 assert(num_se == 1 || num_se == 2 || num_se == 4);
59 assert(sh_per_se == 1 || sh_per_se == 2);
60 assert(rb_per_pkr == 1 || rb_per_pkr == 2);
61
62 /* XXX: I can't figure out what the *_XSEL and *_YSEL
63 * fields are for, so I'm leaving them as their default
64 * values. */
65
66 if ((num_se > 2) && ((!se_mask[0] && !se_mask[1]) ||
67 (!se_mask[2] && !se_mask[3]))) {
68 raster_config_1 &= C_028354_SE_PAIR_MAP;
69
70 if (!se_mask[0] && !se_mask[1]) {
71 raster_config_1 |=
72 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3);
73 } else {
74 raster_config_1 |=
75 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0);
76 }
77 }
78
79 for (se = 0; se < num_se; se++) {
80 unsigned raster_config_se = raster_config;
81 unsigned pkr0_mask = ((1 << rb_per_pkr) - 1) << (se * rb_per_se);
82 unsigned pkr1_mask = pkr0_mask << rb_per_pkr;
83 int idx = (se / 2) * 2;
84
85 if ((num_se > 1) && (!se_mask[idx] || !se_mask[idx + 1])) {
86 raster_config_se &= C_028350_SE_MAP;
87
88 if (!se_mask[idx]) {
89 raster_config_se |=
90 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3);
91 } else {
92 raster_config_se |=
93 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0);
94 }
95 }
96
97 pkr0_mask &= rb_mask;
98 pkr1_mask &= rb_mask;
99 if (rb_per_se > 2 && (!pkr0_mask || !pkr1_mask)) {
100 raster_config_se &= C_028350_PKR_MAP;
101
102 if (!pkr0_mask) {
103 raster_config_se |=
104 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3);
105 } else {
106 raster_config_se |=
107 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0);
108 }
109 }
110
111 if (rb_per_se >= 2) {
112 unsigned rb0_mask = 1 << (se * rb_per_se);
113 unsigned rb1_mask = rb0_mask << 1;
114
115 rb0_mask &= rb_mask;
116 rb1_mask &= rb_mask;
117 if (!rb0_mask || !rb1_mask) {
118 raster_config_se &= C_028350_RB_MAP_PKR0;
119
120 if (!rb0_mask) {
121 raster_config_se |=
122 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3);
123 } else {
124 raster_config_se |=
125 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0);
126 }
127 }
128
129 if (rb_per_se > 2) {
130 rb0_mask = 1 << (se * rb_per_se + rb_per_pkr);
131 rb1_mask = rb0_mask << 1;
132 rb0_mask &= rb_mask;
133 rb1_mask &= rb_mask;
134 if (!rb0_mask || !rb1_mask) {
135 raster_config_se &= C_028350_RB_MAP_PKR1;
136
137 if (!rb0_mask) {
138 raster_config_se |=
139 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3);
140 } else {
141 raster_config_se |=
142 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0);
143 }
144 }
145 }
146 }
147
148 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
149 if (physical_device->rad_info.chip_class < CIK)
150 radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
151 S_00802C_SE_INDEX(se) |
152 S_00802C_SH_BROADCAST_WRITES(1) |
153 S_00802C_INSTANCE_BROADCAST_WRITES(1));
154 else
155 radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
156 S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
157 S_030800_INSTANCE_BROADCAST_WRITES(1));
158 radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG, raster_config_se);
159 if (physical_device->rad_info.chip_class >= CIK)
160 radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
161 }
162
163 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
164 if (physical_device->rad_info.chip_class < CIK)
165 radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
166 S_00802C_SE_BROADCAST_WRITES(1) |
167 S_00802C_SH_BROADCAST_WRITES(1) |
168 S_00802C_INSTANCE_BROADCAST_WRITES(1));
169 else
170 radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
171 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
172 S_030800_INSTANCE_BROADCAST_WRITES(1));
173 }
174
175 static void
176 si_emit_compute(struct radv_physical_device *physical_device,
177 struct radeon_winsys_cs *cs)
178 {
179 radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
180 radeon_emit(cs, 0);
181 radeon_emit(cs, 0);
182 radeon_emit(cs, 0);
183
184 radeon_set_sh_reg_seq(cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
185 S_00B854_WAVES_PER_SH(0x3));
186 radeon_emit(cs, 0);
187 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
188 radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
189 radeon_emit(cs, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
190
191 if (physical_device->rad_info.chip_class >= CIK) {
192 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
193 radeon_set_sh_reg_seq(cs,
194 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
195 radeon_emit(cs, S_00B864_SH0_CU_EN(0xffff) |
196 S_00B864_SH1_CU_EN(0xffff));
197 radeon_emit(cs, S_00B868_SH0_CU_EN(0xffff) |
198 S_00B868_SH1_CU_EN(0xffff));
199 }
200
201 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
202 * and is now per pipe, so it should be handled in the
203 * kernel if we want to use something other than the default value,
204 * which is now 0x22f.
205 */
206 if (physical_device->rad_info.chip_class <= SI) {
207 /* XXX: This should be:
208 * (number of compute units) * 4 * (waves per simd) - 1 */
209
210 radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
211 0x190 /* Default value */);
212 }
213 }
214
215 void
216 si_init_compute(struct radv_cmd_buffer *cmd_buffer)
217 {
218 struct radv_physical_device *physical_device = cmd_buffer->device->physical_device;
219 si_emit_compute(physical_device, cmd_buffer->cs);
220 }
221
222 /* 12.4 fixed-point */
223 static unsigned radv_pack_float_12p4(float x)
224 {
225 return x <= 0 ? 0 :
226 x >= 4096 ? 0xffff : x * 16;
227 }
228
229 static void
230 si_set_raster_config(struct radv_physical_device *physical_device,
231 struct radeon_winsys_cs *cs)
232 {
233 unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
234 unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
235 unsigned raster_config, raster_config_1;
236
237 switch (physical_device->rad_info.family) {
238 case CHIP_TAHITI:
239 case CHIP_PITCAIRN:
240 raster_config = 0x2a00126a;
241 raster_config_1 = 0x00000000;
242 break;
243 case CHIP_VERDE:
244 raster_config = 0x0000124a;
245 raster_config_1 = 0x00000000;
246 break;
247 case CHIP_OLAND:
248 raster_config = 0x00000082;
249 raster_config_1 = 0x00000000;
250 break;
251 case CHIP_HAINAN:
252 raster_config = 0x00000000;
253 raster_config_1 = 0x00000000;
254 break;
255 case CHIP_BONAIRE:
256 raster_config = 0x16000012;
257 raster_config_1 = 0x00000000;
258 break;
259 case CHIP_HAWAII:
260 raster_config = 0x3a00161a;
261 raster_config_1 = 0x0000002e;
262 break;
263 case CHIP_FIJI:
264 if (physical_device->rad_info.cik_macrotile_mode_array[0] == 0x000000e8) {
265 /* old kernels with old tiling config */
266 raster_config = 0x16000012;
267 raster_config_1 = 0x0000002a;
268 } else {
269 raster_config = 0x3a00161a;
270 raster_config_1 = 0x0000002e;
271 }
272 break;
273 case CHIP_POLARIS10:
274 raster_config = 0x16000012;
275 raster_config_1 = 0x0000002a;
276 break;
277 case CHIP_POLARIS11:
278 case CHIP_POLARIS12:
279 raster_config = 0x16000012;
280 raster_config_1 = 0x00000000;
281 break;
282 case CHIP_TONGA:
283 raster_config = 0x16000012;
284 raster_config_1 = 0x0000002a;
285 break;
286 case CHIP_ICELAND:
287 if (num_rb == 1)
288 raster_config = 0x00000000;
289 else
290 raster_config = 0x00000002;
291 raster_config_1 = 0x00000000;
292 break;
293 case CHIP_CARRIZO:
294 raster_config = 0x00000002;
295 raster_config_1 = 0x00000000;
296 break;
297 case CHIP_KAVERI:
298 /* KV should be 0x00000002, but that causes problems with radeon */
299 raster_config = 0x00000000; /* 0x00000002 */
300 raster_config_1 = 0x00000000;
301 break;
302 case CHIP_KABINI:
303 case CHIP_MULLINS:
304 case CHIP_STONEY:
305 raster_config = 0x00000000;
306 raster_config_1 = 0x00000000;
307 break;
308 default:
309 fprintf(stderr,
310 "radv: Unknown GPU, using 0 for raster_config\n");
311 raster_config = 0x00000000;
312 raster_config_1 = 0x00000000;
313 break;
314 }
315
316 /* Always use the default config when all backends are enabled
317 * (or when we failed to determine the enabled backends).
318 */
319 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
320 radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG,
321 raster_config);
322 if (physical_device->rad_info.chip_class >= CIK)
323 radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
324 raster_config_1);
325 } else {
326 si_write_harvested_raster_configs(physical_device, cs,
327 raster_config,
328 raster_config_1);
329 }
330 }
331
332 static void
333 si_emit_config(struct radv_physical_device *physical_device,
334 struct radeon_winsys_cs *cs)
335 {
336 int i;
337
338 /* Only SI can disable CLEAR_STATE for now. */
339 assert(physical_device->has_clear_state ||
340 physical_device->rad_info.chip_class == SI);
341
342 radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
343 radeon_emit(cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
344 radeon_emit(cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
345
346 if (physical_device->has_clear_state) {
347 radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
348 radeon_emit(cs, 0);
349 }
350
351 if (physical_device->rad_info.chip_class <= VI)
352 si_set_raster_config(physical_device, cs);
353
354 radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
355 if (!physical_device->has_clear_state)
356 radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
357
358 /* FIXME calculate these values somehow ??? */
359 if (physical_device->rad_info.chip_class <= VI) {
360 radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
361 radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
362 }
363
364 if (!physical_device->has_clear_state) {
365 radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
366 radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
367 radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
368 }
369
370 radeon_set_context_reg(cs, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
371 if (!physical_device->has_clear_state)
372 radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
373 if (physical_device->rad_info.chip_class < CIK)
374 radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
375 S_008A14_CLIP_VTX_REORDER_ENA(1));
376
377 radeon_set_context_reg(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
378 radeon_set_context_reg(cs, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
379
380 if (!physical_device->has_clear_state)
381 radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
382
383 /* CLEAR_STATE doesn't clear these correctly on certain generations.
384 * I don't know why. Deduced by trial and error.
385 */
386 if (physical_device->rad_info.chip_class <= CIK) {
387 radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
388 radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL,
389 S_028204_WINDOW_OFFSET_DISABLE(1));
390 radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL,
391 S_028240_WINDOW_OFFSET_DISABLE(1));
392 radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
393 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
394 radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
395 radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
396 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
397 }
398
399 if (!physical_device->has_clear_state) {
400 for (i = 0; i < 16; i++) {
401 radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
402 radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
403 }
404 }
405
406 if (!physical_device->has_clear_state) {
407 radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
408 radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
409 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
410 radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
411 radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
412 radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
413 radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
414 radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
415 }
416
417 radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE,
418 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
419 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
420
421 if (physical_device->rad_info.chip_class >= GFX9) {
422 radeon_set_uconfig_reg(cs, R_030920_VGT_MAX_VTX_INDX, ~0);
423 radeon_set_uconfig_reg(cs, R_030924_VGT_MIN_VTX_INDX, 0);
424 radeon_set_uconfig_reg(cs, R_030928_VGT_INDX_OFFSET, 0);
425 } else {
426 /* These registers, when written, also overwrite the
427 * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
428 * them. It would be an issue if there was another UMD
429 * changing them.
430 */
431 radeon_set_context_reg(cs, R_028400_VGT_MAX_VTX_INDX, ~0);
432 radeon_set_context_reg(cs, R_028404_VGT_MIN_VTX_INDX, 0);
433 radeon_set_context_reg(cs, R_028408_VGT_INDX_OFFSET, 0);
434 }
435
436 if (physical_device->rad_info.chip_class >= CIK) {
437 if (physical_device->rad_info.chip_class >= GFX9) {
438 radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
439 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
440 } else {
441 radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
442 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
443 radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
444 S_00B41C_WAVE_LIMIT(0x3F));
445 radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
446 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
447 /* If this is 0, Bonaire can hang even if GS isn't being used.
448 * Other chips are unaffected. These are suboptimal values,
449 * but we don't use on-chip GS.
450 */
451 radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL,
452 S_028A44_ES_VERTS_PER_SUBGRP(64) |
453 S_028A44_GS_PRIMS_PER_SUBGRP(4));
454 }
455 radeon_set_sh_reg(cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
456 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
457
458 if (physical_device->rad_info.num_good_compute_units /
459 (physical_device->rad_info.max_se * physical_device->rad_info.max_sh_per_se) <= 4) {
460 /* Too few available compute units per SH. Disallowing
461 * VS to run on CU0 could hurt us more than late VS
462 * allocation would help.
463 *
464 * LATE_ALLOC_VS = 2 is the highest safe number.
465 */
466 radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
467 S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
468 radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
469 } else {
470 /* Set LATE_ALLOC_VS == 31. It should be less than
471 * the number of scratch waves. Limitations:
472 * - VS can't execute on CU0.
473 * - If HS writes outputs to LDS, LS can't execute on CU0.
474 */
475 radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
476 S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
477 radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
478 }
479
480 radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
481 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
482 }
483
484 if (physical_device->rad_info.chip_class >= VI) {
485 uint32_t vgt_tess_distribution;
486 radeon_set_context_reg(cs, R_028424_CB_DCC_CONTROL,
487 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
488 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
489
490 vgt_tess_distribution = S_028B50_ACCUM_ISOLINE(32) |
491 S_028B50_ACCUM_TRI(11) |
492 S_028B50_ACCUM_QUAD(11) |
493 S_028B50_DONUT_SPLIT(16);
494
495 if (physical_device->rad_info.family == CHIP_FIJI ||
496 physical_device->rad_info.family >= CHIP_POLARIS10)
497 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
498
499 radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
500 vgt_tess_distribution);
501 } else if (!physical_device->has_clear_state) {
502 radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
503 radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
504 }
505
506 if (physical_device->rad_info.chip_class >= GFX9) {
507 unsigned num_se = physical_device->rad_info.max_se;
508 unsigned pc_lines = 0;
509
510 switch (physical_device->rad_info.family) {
511 case CHIP_VEGA10:
512 pc_lines = 4096;
513 break;
514 case CHIP_RAVEN:
515 pc_lines = 1024;
516 break;
517 default:
518 assert(0);
519 }
520
521 radeon_set_context_reg(cs, R_028060_DB_DFSM_CONTROL,
522 S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF));
523 /* TODO: Enable the binner: */
524 radeon_set_context_reg(cs, R_028C44_PA_SC_BINNER_CNTL_0,
525 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
526 S_028C44_DISABLE_START_OF_PRIM(1));
527 radeon_set_context_reg(cs, R_028C48_PA_SC_BINNER_CNTL_1,
528 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
529 S_028C48_MAX_PRIM_PER_BATCH(1023));
530 radeon_set_context_reg(cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
531 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
532 radeon_set_uconfig_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
533 }
534
535 unsigned tmp = (unsigned)(1.0 * 8.0);
536 radeon_set_context_reg_seq(cs, R_028A00_PA_SU_POINT_SIZE, 1);
537 radeon_emit(cs, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
538 radeon_set_context_reg_seq(cs, R_028A04_PA_SU_POINT_MINMAX, 1);
539 radeon_emit(cs, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
540 S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
541
542 if (!physical_device->has_clear_state) {
543 radeon_set_context_reg(cs, R_028004_DB_COUNT_CONTROL,
544 S_028004_ZPASS_INCREMENT_DISABLE(1));
545 }
546
547 si_emit_compute(physical_device, cs);
548 }
549
550 void si_init_config(struct radv_cmd_buffer *cmd_buffer)
551 {
552 struct radv_physical_device *physical_device = cmd_buffer->device->physical_device;
553
554 si_emit_config(physical_device, cmd_buffer->cs);
555 }
556
557 void
558 cik_create_gfx_config(struct radv_device *device)
559 {
560 struct radeon_winsys_cs *cs = device->ws->cs_create(device->ws, RING_GFX);
561 if (!cs)
562 return;
563
564 si_emit_config(device->physical_device, cs);
565
566 while (cs->cdw & 7) {
567 if (device->physical_device->rad_info.gfx_ib_pad_with_type2)
568 radeon_emit(cs, 0x80000000);
569 else
570 radeon_emit(cs, 0xffff1000);
571 }
572
573 device->gfx_init = device->ws->buffer_create(device->ws,
574 cs->cdw * 4, 4096,
575 RADEON_DOMAIN_GTT,
576 RADEON_FLAG_CPU_ACCESS|
577 RADEON_FLAG_NO_INTERPROCESS_SHARING);
578 if (!device->gfx_init)
579 goto fail;
580
581 void *map = device->ws->buffer_map(device->gfx_init);
582 if (!map) {
583 device->ws->buffer_destroy(device->gfx_init);
584 device->gfx_init = NULL;
585 goto fail;
586 }
587 memcpy(map, cs->buf, cs->cdw * 4);
588
589 device->ws->buffer_unmap(device->gfx_init);
590 device->gfx_init_size_dw = cs->cdw;
591 fail:
592 device->ws->cs_destroy(cs);
593 }
594
595 static void
596 get_viewport_xform(const VkViewport *viewport,
597 float scale[3], float translate[3])
598 {
599 float x = viewport->x;
600 float y = viewport->y;
601 float half_width = 0.5f * viewport->width;
602 float half_height = 0.5f * viewport->height;
603 double n = viewport->minDepth;
604 double f = viewport->maxDepth;
605
606 scale[0] = half_width;
607 translate[0] = half_width + x;
608 scale[1] = half_height;
609 translate[1] = half_height + y;
610
611 scale[2] = (f - n);
612 translate[2] = n;
613 }
614
615 void
616 si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
617 int count, const VkViewport *viewports)
618 {
619 int i;
620
621 assert(count);
622 radeon_set_context_reg_seq(cs, R_02843C_PA_CL_VPORT_XSCALE +
623 first_vp * 4 * 6, count * 6);
624
625 for (i = 0; i < count; i++) {
626 float scale[3], translate[3];
627
628
629 get_viewport_xform(&viewports[i], scale, translate);
630 radeon_emit(cs, fui(scale[0]));
631 radeon_emit(cs, fui(translate[0]));
632 radeon_emit(cs, fui(scale[1]));
633 radeon_emit(cs, fui(translate[1]));
634 radeon_emit(cs, fui(scale[2]));
635 radeon_emit(cs, fui(translate[2]));
636 }
637
638 radeon_set_context_reg_seq(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 +
639 first_vp * 4 * 2, count * 2);
640 for (i = 0; i < count; i++) {
641 float zmin = MIN2(viewports[i].minDepth, viewports[i].maxDepth);
642 float zmax = MAX2(viewports[i].minDepth, viewports[i].maxDepth);
643 radeon_emit(cs, fui(zmin));
644 radeon_emit(cs, fui(zmax));
645 }
646 }
647
648 static VkRect2D si_scissor_from_viewport(const VkViewport *viewport)
649 {
650 float scale[3], translate[3];
651 VkRect2D rect;
652
653 get_viewport_xform(viewport, scale, translate);
654
655 rect.offset.x = translate[0] - abs(scale[0]);
656 rect.offset.y = translate[1] - abs(scale[1]);
657 rect.extent.width = ceilf(translate[0] + abs(scale[0])) - rect.offset.x;
658 rect.extent.height = ceilf(translate[1] + abs(scale[1])) - rect.offset.y;
659
660 return rect;
661 }
662
663 static VkRect2D si_intersect_scissor(const VkRect2D *a, const VkRect2D *b) {
664 VkRect2D ret;
665 ret.offset.x = MAX2(a->offset.x, b->offset.x);
666 ret.offset.y = MAX2(a->offset.y, b->offset.y);
667 ret.extent.width = MIN2(a->offset.x + a->extent.width,
668 b->offset.x + b->extent.width) - ret.offset.x;
669 ret.extent.height = MIN2(a->offset.y + a->extent.height,
670 b->offset.y + b->extent.height) - ret.offset.y;
671 return ret;
672 }
673
674 void
675 si_write_scissors(struct radeon_winsys_cs *cs, int first,
676 int count, const VkRect2D *scissors,
677 const VkViewport *viewports, bool can_use_guardband)
678 {
679 int i;
680 float scale[3], translate[3], guardband_x = INFINITY, guardband_y = INFINITY;
681 const float max_range = 32767.0f;
682 assert(count);
683
684 radeon_set_context_reg_seq(cs, R_028250_PA_SC_VPORT_SCISSOR_0_TL + first * 4 * 2, count * 2);
685 for (i = 0; i < count; i++) {
686 VkRect2D viewport_scissor = si_scissor_from_viewport(viewports + i);
687 VkRect2D scissor = si_intersect_scissor(&scissors[i], &viewport_scissor);
688
689 get_viewport_xform(viewports + i, scale, translate);
690 scale[0] = abs(scale[0]);
691 scale[1] = abs(scale[1]);
692
693 if (scale[0] < 0.5)
694 scale[0] = 0.5;
695 if (scale[1] < 0.5)
696 scale[1] = 0.5;
697
698 guardband_x = MIN2(guardband_x, (max_range - abs(translate[0])) / scale[0]);
699 guardband_y = MIN2(guardband_y, (max_range - abs(translate[1])) / scale[1]);
700
701 radeon_emit(cs, S_028250_TL_X(scissor.offset.x) |
702 S_028250_TL_Y(scissor.offset.y) |
703 S_028250_WINDOW_OFFSET_DISABLE(1));
704 radeon_emit(cs, S_028254_BR_X(scissor.offset.x + scissor.extent.width) |
705 S_028254_BR_Y(scissor.offset.y + scissor.extent.height));
706 }
707 if (!can_use_guardband) {
708 guardband_x = 1.0;
709 guardband_y = 1.0;
710 }
711
712 radeon_set_context_reg_seq(cs, R_028BE8_PA_CL_GB_VERT_CLIP_ADJ, 4);
713 radeon_emit(cs, fui(guardband_y));
714 radeon_emit(cs, fui(1.0));
715 radeon_emit(cs, fui(guardband_x));
716 radeon_emit(cs, fui(1.0));
717 }
718
719 static inline unsigned
720 radv_prims_for_vertices(struct radv_prim_vertex_count *info, unsigned num)
721 {
722 if (num == 0)
723 return 0;
724
725 if (info->incr == 0)
726 return 0;
727
728 if (num < info->min)
729 return 0;
730
731 return 1 + ((num - info->min) / info->incr);
732 }
733
734 uint32_t
735 si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
736 bool instanced_draw, bool indirect_draw,
737 uint32_t draw_vertex_count)
738 {
739 enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
740 enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
741 struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
742 const unsigned max_primgroup_in_wave = 2;
743 /* SWITCH_ON_EOP(0) is always preferable. */
744 bool wd_switch_on_eop = false;
745 bool ia_switch_on_eop = false;
746 bool ia_switch_on_eoi = false;
747 bool partial_vs_wave = false;
748 bool partial_es_wave = cmd_buffer->state.pipeline->graphics.partial_es_wave;
749 bool multi_instances_smaller_than_primgroup;
750
751 multi_instances_smaller_than_primgroup = indirect_draw;
752 if (!multi_instances_smaller_than_primgroup && instanced_draw) {
753 uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
754 if (num_prims < cmd_buffer->state.pipeline->graphics.primgroup_size)
755 multi_instances_smaller_than_primgroup = true;
756 }
757
758 ia_switch_on_eoi = cmd_buffer->state.pipeline->graphics.ia_switch_on_eoi;
759 partial_vs_wave = cmd_buffer->state.pipeline->graphics.partial_vs_wave;
760
761 if (chip_class >= CIK) {
762 wd_switch_on_eop = cmd_buffer->state.pipeline->graphics.wd_switch_on_eop;
763
764 /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
765 * We don't know that for indirect drawing, so treat it as
766 * always problematic. */
767 if (family == CHIP_HAWAII &&
768 (instanced_draw || indirect_draw))
769 wd_switch_on_eop = true;
770
771 /* Performance recommendation for 4 SE Gfx7-8 parts if
772 * instances are smaller than a primgroup.
773 * Assume indirect draws always use small instances.
774 * This is needed for good VS wave utilization.
775 */
776 if (chip_class <= VI &&
777 info->max_se == 4 &&
778 multi_instances_smaller_than_primgroup)
779 wd_switch_on_eop = true;
780
781 /* Required on CIK and later. */
782 if (info->max_se > 2 && !wd_switch_on_eop)
783 ia_switch_on_eoi = true;
784
785 /* Required by Hawaii and, for some special cases, by VI. */
786 if (ia_switch_on_eoi &&
787 (family == CHIP_HAWAII ||
788 (chip_class == VI &&
789 /* max primgroup in wave is always 2 - leave this for documentation */
790 (radv_pipeline_has_gs(cmd_buffer->state.pipeline) || max_primgroup_in_wave != 2))))
791 partial_vs_wave = true;
792
793 /* Instancing bug on Bonaire. */
794 if (family == CHIP_BONAIRE && ia_switch_on_eoi &&
795 (instanced_draw || indirect_draw))
796 partial_vs_wave = true;
797
798 /* If the WD switch is false, the IA switch must be false too. */
799 assert(wd_switch_on_eop || !ia_switch_on_eop);
800 }
801 /* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
802 if (chip_class <= VI && ia_switch_on_eoi)
803 partial_es_wave = true;
804
805 if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
806 /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
807 * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
808 * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
809 */
810 if (family == CHIP_HAWAII && ia_switch_on_eoi) {
811 bool set_vgt_flush = indirect_draw;
812 if (!set_vgt_flush && instanced_draw) {
813 uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
814 if (num_prims <= 1)
815 set_vgt_flush = true;
816 }
817 if (set_vgt_flush)
818 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VGT_FLUSH;
819 }
820 }
821
822 return cmd_buffer->state.pipeline->graphics.base_ia_multi_vgt_param |
823 S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
824 S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
825 S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
826 S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
827 S_028AA8_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0);
828
829 }
830
831 void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
832 bool predicated,
833 enum chip_class chip_class,
834 bool is_mec,
835 unsigned event, unsigned event_flags,
836 unsigned data_sel,
837 uint64_t va,
838 uint32_t old_fence,
839 uint32_t new_fence)
840 {
841 unsigned op = EVENT_TYPE(event) |
842 EVENT_INDEX(5) |
843 event_flags;
844 unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
845
846 if (chip_class >= GFX9 || is_gfx8_mec) {
847 radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, predicated));
848 radeon_emit(cs, op);
849 radeon_emit(cs, EOP_DATA_SEL(data_sel));
850 radeon_emit(cs, va); /* address lo */
851 radeon_emit(cs, va >> 32); /* address hi */
852 radeon_emit(cs, new_fence); /* immediate data lo */
853 radeon_emit(cs, 0); /* immediate data hi */
854 if (!is_gfx8_mec)
855 radeon_emit(cs, 0); /* unused */
856 } else {
857 if (chip_class == CIK ||
858 chip_class == VI) {
859 /* Two EOP events are required to make all engines go idle
860 * (and optional cache flushes executed) before the timestamp
861 * is written.
862 */
863 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
864 radeon_emit(cs, op);
865 radeon_emit(cs, va);
866 radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
867 radeon_emit(cs, old_fence); /* immediate data */
868 radeon_emit(cs, 0); /* unused */
869 }
870
871 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
872 radeon_emit(cs, op);
873 radeon_emit(cs, va);
874 radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
875 radeon_emit(cs, new_fence); /* immediate data */
876 radeon_emit(cs, 0); /* unused */
877 }
878 }
879
880 void
881 si_emit_wait_fence(struct radeon_winsys_cs *cs,
882 bool predicated,
883 uint64_t va, uint32_t ref,
884 uint32_t mask)
885 {
886 radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, predicated));
887 radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
888 radeon_emit(cs, va);
889 radeon_emit(cs, va >> 32);
890 radeon_emit(cs, ref); /* reference value */
891 radeon_emit(cs, mask); /* mask */
892 radeon_emit(cs, 4); /* poll interval */
893 }
894
895 static void
896 si_emit_acquire_mem(struct radeon_winsys_cs *cs,
897 bool is_mec,
898 bool predicated,
899 bool is_gfx9,
900 unsigned cp_coher_cntl)
901 {
902 if (is_mec || is_gfx9) {
903 uint32_t hi_val = is_gfx9 ? 0xffffff : 0xff;
904 radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, predicated) |
905 PKT3_SHADER_TYPE_S(is_mec));
906 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
907 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
908 radeon_emit(cs, hi_val); /* CP_COHER_SIZE_HI */
909 radeon_emit(cs, 0); /* CP_COHER_BASE */
910 radeon_emit(cs, 0); /* CP_COHER_BASE_HI */
911 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
912 } else {
913 /* ACQUIRE_MEM is only required on a compute ring. */
914 radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, predicated));
915 radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
916 radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
917 radeon_emit(cs, 0); /* CP_COHER_BASE */
918 radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
919 }
920 }
921
922 void
923 si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
924 bool predicated,
925 enum chip_class chip_class,
926 uint32_t *flush_cnt,
927 uint64_t flush_va,
928 bool is_mec,
929 enum radv_cmd_flush_bits flush_bits)
930 {
931 unsigned cp_coher_cntl = 0;
932 uint32_t flush_cb_db = flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
933 RADV_CMD_FLAG_FLUSH_AND_INV_DB);
934
935 if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
936 cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
937 if (flush_bits & RADV_CMD_FLAG_INV_SMEM_L1)
938 cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
939
940 if (chip_class <= VI) {
941 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
942 cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
943 S_0085F0_CB0_DEST_BASE_ENA(1) |
944 S_0085F0_CB1_DEST_BASE_ENA(1) |
945 S_0085F0_CB2_DEST_BASE_ENA(1) |
946 S_0085F0_CB3_DEST_BASE_ENA(1) |
947 S_0085F0_CB4_DEST_BASE_ENA(1) |
948 S_0085F0_CB5_DEST_BASE_ENA(1) |
949 S_0085F0_CB6_DEST_BASE_ENA(1) |
950 S_0085F0_CB7_DEST_BASE_ENA(1);
951
952 /* Necessary for DCC */
953 if (chip_class >= VI) {
954 si_cs_emit_write_event_eop(cs,
955 predicated,
956 chip_class,
957 is_mec,
958 V_028A90_FLUSH_AND_INV_CB_DATA_TS,
959 0, 0, 0, 0, 0);
960 }
961 }
962 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
963 cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
964 S_0085F0_DB_DEST_BASE_ENA(1);
965 }
966 }
967
968 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) {
969 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
970 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
971 }
972
973 if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) {
974 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
975 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
976 }
977
978 if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
979 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
980 radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
981 } else if (flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
982 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
983 radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
984 }
985
986 if (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
987 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
988 radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
989 }
990
991 if (chip_class >= GFX9 && flush_cb_db) {
992 unsigned cb_db_event, tc_flags;
993
994 #if 0
995 /* This breaks a bunch of:
996 dEQP-VK.renderpass.dedicated_allocation.formats.d32_sfloat_s8_uint.input*.
997 use the big hammer always.
998 */
999 /* Set the CB/DB flush event. */
1000 switch (flush_cb_db) {
1001 case RADV_CMD_FLAG_FLUSH_AND_INV_CB:
1002 cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
1003 break;
1004 case RADV_CMD_FLAG_FLUSH_AND_INV_DB:
1005 cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
1006 break;
1007 default:
1008 /* both CB & DB */
1009 cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
1010 }
1011 #else
1012 cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
1013 #endif
1014 /* TC | TC_WB = invalidate L2 data
1015 * TC_MD | TC_WB = invalidate L2 metadata
1016 * TC | TC_WB | TC_MD = invalidate L2 data & metadata
1017 *
1018 * The metadata cache must always be invalidated for coherency
1019 * between CB/DB and shaders. (metadata = HTILE, CMASK, DCC)
1020 *
1021 * TC must be invalidated on GFX9 only if the CB/DB surface is
1022 * not pipe-aligned. If the surface is RB-aligned, it might not
1023 * strictly be pipe-aligned since RB alignment takes precendence.
1024 */
1025 tc_flags = EVENT_TC_WB_ACTION_ENA |
1026 EVENT_TC_MD_ACTION_ENA;
1027
1028 /* Ideally flush TC together with CB/DB. */
1029 if (flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) {
1030 tc_flags |= EVENT_TC_ACTION_ENA |
1031 EVENT_TCL1_ACTION_ENA;
1032
1033 /* Clear the flags. */
1034 flush_bits &= ~(RADV_CMD_FLAG_INV_GLOBAL_L2 |
1035 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 |
1036 RADV_CMD_FLAG_INV_VMEM_L1);
1037 }
1038 assert(flush_cnt);
1039 uint32_t old_fence = (*flush_cnt)++;
1040
1041 si_cs_emit_write_event_eop(cs, predicated, chip_class, false, cb_db_event, tc_flags, 1,
1042 flush_va, old_fence, *flush_cnt);
1043 si_emit_wait_fence(cs, predicated, flush_va, *flush_cnt, 0xffffffff);
1044 }
1045
1046 /* VGT state sync */
1047 if (flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
1048 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
1049 radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1050 }
1051
1052 /* Make sure ME is idle (it executes most packets) before continuing.
1053 * This prevents read-after-write hazards between PFP and ME.
1054 */
1055 if ((cp_coher_cntl ||
1056 (flush_bits & (RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
1057 RADV_CMD_FLAG_INV_VMEM_L1 |
1058 RADV_CMD_FLAG_INV_GLOBAL_L2 |
1059 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) &&
1060 !is_mec) {
1061 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, predicated));
1062 radeon_emit(cs, 0);
1063 }
1064
1065 if ((flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) ||
1066 (chip_class <= CIK && (flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) {
1067 si_emit_acquire_mem(cs, is_mec, predicated, chip_class >= GFX9,
1068 cp_coher_cntl |
1069 S_0085F0_TC_ACTION_ENA(1) |
1070 S_0085F0_TCL1_ACTION_ENA(1) |
1071 S_0301F0_TC_WB_ACTION_ENA(chip_class >= VI));
1072 cp_coher_cntl = 0;
1073 } else {
1074 if(flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2) {
1075 /* WB = write-back
1076 * NC = apply to non-coherent MTYPEs
1077 * (i.e. MTYPE <= 1, which is what we use everywhere)
1078 *
1079 * WB doesn't work without NC.
1080 */
1081 si_emit_acquire_mem(cs, is_mec, predicated,
1082 chip_class >= GFX9,
1083 cp_coher_cntl |
1084 S_0301F0_TC_WB_ACTION_ENA(1) |
1085 S_0301F0_TC_NC_ACTION_ENA(1));
1086 cp_coher_cntl = 0;
1087 }
1088 if (flush_bits & RADV_CMD_FLAG_INV_VMEM_L1) {
1089 si_emit_acquire_mem(cs, is_mec,
1090 predicated, chip_class >= GFX9,
1091 cp_coher_cntl |
1092 S_0085F0_TCL1_ACTION_ENA(1));
1093 cp_coher_cntl = 0;
1094 }
1095 }
1096
1097 /* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
1098 * Therefore, it should be last. Done in PFP.
1099 */
1100 if (cp_coher_cntl)
1101 si_emit_acquire_mem(cs, is_mec, predicated, chip_class >= GFX9, cp_coher_cntl);
1102 }
1103
1104 void
1105 si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
1106 {
1107 bool is_compute = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
1108
1109 if (is_compute)
1110 cmd_buffer->state.flush_bits &= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1111 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
1112 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1113 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
1114 RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
1115 RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
1116 RADV_CMD_FLAG_VGT_FLUSH);
1117
1118 if (!cmd_buffer->state.flush_bits)
1119 return;
1120
1121 enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
1122 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 128);
1123
1124 uint32_t *ptr = NULL;
1125 uint64_t va = 0;
1126 if (chip_class == GFX9) {
1127 va = radv_buffer_get_va(cmd_buffer->gfx9_fence_bo) + cmd_buffer->gfx9_fence_offset;
1128 ptr = &cmd_buffer->gfx9_fence_idx;
1129 }
1130 si_cs_emit_cache_flush(cmd_buffer->cs,
1131 cmd_buffer->state.predicating,
1132 cmd_buffer->device->physical_device->rad_info.chip_class,
1133 ptr, va,
1134 radv_cmd_buffer_uses_mec(cmd_buffer),
1135 cmd_buffer->state.flush_bits);
1136
1137
1138 if (unlikely(cmd_buffer->device->trace_bo))
1139 radv_cmd_buffer_trace_emit(cmd_buffer);
1140
1141 cmd_buffer->state.flush_bits = 0;
1142 }
1143
1144 /* sets the CP predication state using a boolean stored at va */
1145 void
1146 si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va)
1147 {
1148 uint32_t op = 0;
1149
1150 if (va)
1151 op = PRED_OP(PREDICATION_OP_BOOL64) | PREDICATION_DRAW_VISIBLE;
1152 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
1153 radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
1154 radeon_emit(cmd_buffer->cs, op);
1155 radeon_emit(cmd_buffer->cs, va);
1156 radeon_emit(cmd_buffer->cs, va >> 32);
1157 } else {
1158 radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
1159 radeon_emit(cmd_buffer->cs, va);
1160 radeon_emit(cmd_buffer->cs, op | ((va >> 32) & 0xFF));
1161 }
1162 }
1163
1164 /* Set this if you want the 3D engine to wait until CP DMA is done.
1165 * It should be set on the last CP DMA packet. */
1166 #define CP_DMA_SYNC (1 << 0)
1167
1168 /* Set this if the source data was used as a destination in a previous CP DMA
1169 * packet. It's for preventing a read-after-write (RAW) hazard between two
1170 * CP DMA packets. */
1171 #define CP_DMA_RAW_WAIT (1 << 1)
1172 #define CP_DMA_USE_L2 (1 << 2)
1173 #define CP_DMA_CLEAR (1 << 3)
1174
1175 /* Alignment for optimal performance. */
1176 #define SI_CPDMA_ALIGNMENT 32
1177
1178 /* The max number of bytes that can be copied per packet. */
1179 static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer *cmd_buffer)
1180 {
1181 unsigned max = cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 ?
1182 S_414_BYTE_COUNT_GFX9(~0u) :
1183 S_414_BYTE_COUNT_GFX6(~0u);
1184
1185 /* make it aligned for optimal performance */
1186 return max & ~(SI_CPDMA_ALIGNMENT - 1);
1187 }
1188
1189 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
1190 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
1191 * clear value.
1192 */
1193 static void si_emit_cp_dma(struct radv_cmd_buffer *cmd_buffer,
1194 uint64_t dst_va, uint64_t src_va,
1195 unsigned size, unsigned flags)
1196 {
1197 struct radeon_winsys_cs *cs = cmd_buffer->cs;
1198 uint32_t header = 0, command = 0;
1199
1200 assert(size);
1201 assert(size <= cp_dma_max_byte_count(cmd_buffer));
1202
1203 radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
1204 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1205 command |= S_414_BYTE_COUNT_GFX9(size);
1206 else
1207 command |= S_414_BYTE_COUNT_GFX6(size);
1208
1209 /* Sync flags. */
1210 if (flags & CP_DMA_SYNC)
1211 header |= S_411_CP_SYNC(1);
1212 else {
1213 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
1214 command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
1215 else
1216 command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
1217 }
1218
1219 if (flags & CP_DMA_RAW_WAIT)
1220 command |= S_414_RAW_WAIT(1);
1221
1222 /* Src and dst flags. */
1223 if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 &&
1224 !(flags & CP_DMA_CLEAR) &&
1225 src_va == dst_va)
1226 header |= S_411_DSL_SEL(V_411_NOWHERE); /* prefetch only */
1227 else if (flags & CP_DMA_USE_L2)
1228 header |= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2);
1229
1230 if (flags & CP_DMA_CLEAR)
1231 header |= S_411_SRC_SEL(V_411_DATA);
1232 else if (flags & CP_DMA_USE_L2)
1233 header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
1234
1235 if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
1236 radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, cmd_buffer->state.predicating));
1237 radeon_emit(cs, header);
1238 radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
1239 radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
1240 radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
1241 radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
1242 radeon_emit(cs, command);
1243 } else {
1244 assert(!(flags & CP_DMA_USE_L2));
1245 header |= S_411_SRC_ADDR_HI(src_va >> 32);
1246 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, cmd_buffer->state.predicating));
1247 radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
1248 radeon_emit(cs, header); /* SRC_ADDR_HI [15:0] + flags. */
1249 radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
1250 radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
1251 radeon_emit(cs, command);
1252 }
1253
1254 /* CP DMA is executed in ME, but index buffers are read by PFP.
1255 * This ensures that ME (CP DMA) is idle before PFP starts fetching
1256 * indices. If we wanted to execute CP DMA in PFP, this packet
1257 * should precede it.
1258 */
1259 if ((flags & CP_DMA_SYNC) && cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL) {
1260 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
1261 radeon_emit(cs, 0);
1262 }
1263
1264 if (unlikely(cmd_buffer->device->trace_bo))
1265 radv_cmd_buffer_trace_emit(cmd_buffer);
1266 }
1267
1268 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1269 unsigned size)
1270 {
1271 uint64_t aligned_va = va & ~(SI_CPDMA_ALIGNMENT - 1);
1272 uint64_t aligned_size = ((va + size + SI_CPDMA_ALIGNMENT -1) & ~(SI_CPDMA_ALIGNMENT - 1)) - aligned_va;
1273
1274 si_emit_cp_dma(cmd_buffer, aligned_va, aligned_va,
1275 aligned_size, CP_DMA_USE_L2);
1276 }
1277
1278 static void si_cp_dma_prepare(struct radv_cmd_buffer *cmd_buffer, uint64_t byte_count,
1279 uint64_t remaining_size, unsigned *flags)
1280 {
1281
1282 /* Flush the caches for the first copy only.
1283 * Also wait for the previous CP DMA operations.
1284 */
1285 if (cmd_buffer->state.flush_bits) {
1286 si_emit_cache_flush(cmd_buffer);
1287 *flags |= CP_DMA_RAW_WAIT;
1288 }
1289
1290 /* Do the synchronization after the last dma, so that all data
1291 * is written to memory.
1292 */
1293 if (byte_count == remaining_size)
1294 *flags |= CP_DMA_SYNC;
1295 }
1296
1297 static void si_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigned size)
1298 {
1299 uint64_t va;
1300 uint32_t offset;
1301 unsigned dma_flags = 0;
1302 unsigned buf_size = SI_CPDMA_ALIGNMENT * 2;
1303 void *ptr;
1304
1305 assert(size < SI_CPDMA_ALIGNMENT);
1306
1307 radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, SI_CPDMA_ALIGNMENT, &offset, &ptr);
1308
1309 va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
1310 va += offset;
1311
1312 si_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
1313
1314 si_emit_cp_dma(cmd_buffer, va, va + SI_CPDMA_ALIGNMENT, size,
1315 dma_flags);
1316 }
1317
1318 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
1319 uint64_t src_va, uint64_t dest_va,
1320 uint64_t size)
1321 {
1322 uint64_t main_src_va, main_dest_va;
1323 uint64_t skipped_size = 0, realign_size = 0;
1324
1325
1326 if (cmd_buffer->device->physical_device->rad_info.family <= CHIP_CARRIZO ||
1327 cmd_buffer->device->physical_device->rad_info.family == CHIP_STONEY) {
1328 /* If the size is not aligned, we must add a dummy copy at the end
1329 * just to align the internal counter. Otherwise, the DMA engine
1330 * would slow down by an order of magnitude for following copies.
1331 */
1332 if (size % SI_CPDMA_ALIGNMENT)
1333 realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
1334
1335 /* If the copy begins unaligned, we must start copying from the next
1336 * aligned block and the skipped part should be copied after everything
1337 * else has been copied. Only the src alignment matters, not dst.
1338 */
1339 if (src_va % SI_CPDMA_ALIGNMENT) {
1340 skipped_size = SI_CPDMA_ALIGNMENT - (src_va % SI_CPDMA_ALIGNMENT);
1341 /* The main part will be skipped if the size is too small. */
1342 skipped_size = MIN2(skipped_size, size);
1343 size -= skipped_size;
1344 }
1345 }
1346 main_src_va = src_va + skipped_size;
1347 main_dest_va = dest_va + skipped_size;
1348
1349 while (size) {
1350 unsigned dma_flags = 0;
1351 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1352
1353 si_cp_dma_prepare(cmd_buffer, byte_count,
1354 size + skipped_size + realign_size,
1355 &dma_flags);
1356
1357 si_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va,
1358 byte_count, dma_flags);
1359
1360 size -= byte_count;
1361 main_src_va += byte_count;
1362 main_dest_va += byte_count;
1363 }
1364
1365 if (skipped_size) {
1366 unsigned dma_flags = 0;
1367
1368 si_cp_dma_prepare(cmd_buffer, skipped_size,
1369 size + skipped_size + realign_size,
1370 &dma_flags);
1371
1372 si_emit_cp_dma(cmd_buffer, dest_va, src_va,
1373 skipped_size, dma_flags);
1374 }
1375 if (realign_size)
1376 si_cp_dma_realign_engine(cmd_buffer, realign_size);
1377 }
1378
1379 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1380 uint64_t size, unsigned value)
1381 {
1382
1383 if (!size)
1384 return;
1385
1386 assert(va % 4 == 0 && size % 4 == 0);
1387
1388 while (size) {
1389 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
1390 unsigned dma_flags = CP_DMA_CLEAR;
1391
1392 si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
1393
1394 /* Emit the clear packet. */
1395 si_emit_cp_dma(cmd_buffer, va, value, byte_count,
1396 dma_flags);
1397
1398 size -= byte_count;
1399 va += byte_count;
1400 }
1401 }
1402
1403 /* For MSAA sample positions. */
1404 #define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
1405 (((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) | \
1406 (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
1407 (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
1408 (((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
1409
1410
1411 /* 2xMSAA
1412 * There are two locations (4, 4), (-4, -4). */
1413 const uint32_t eg_sample_locs_2x[4] = {
1414 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1415 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1416 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1417 FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
1418 };
1419 const unsigned eg_max_dist_2x = 4;
1420 /* 4xMSAA
1421 * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
1422 const uint32_t eg_sample_locs_4x[4] = {
1423 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1424 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1425 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1426 FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
1427 };
1428 const unsigned eg_max_dist_4x = 6;
1429
1430 /* Cayman 8xMSAA */
1431 static const uint32_t cm_sample_locs_8x[] = {
1432 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1433 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1434 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1435 FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
1436 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1437 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1438 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1439 FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
1440 };
1441 static const unsigned cm_max_dist_8x = 8;
1442 /* Cayman 16xMSAA */
1443 static const uint32_t cm_sample_locs_16x[] = {
1444 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1445 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1446 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1447 FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
1448 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1449 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1450 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1451 FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
1452 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1453 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1454 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1455 FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
1456 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1457 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1458 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1459 FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
1460 };
1461 static const unsigned cm_max_dist_16x = 8;
1462
1463 unsigned radv_cayman_get_maxdist(int log_samples)
1464 {
1465 unsigned max_dist[] = {
1466 0,
1467 eg_max_dist_2x,
1468 eg_max_dist_4x,
1469 cm_max_dist_8x,
1470 cm_max_dist_16x
1471 };
1472 return max_dist[log_samples];
1473 }
1474
1475 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples)
1476 {
1477 switch (nr_samples) {
1478 default:
1479 case 1:
1480 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 0);
1481 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, 0);
1482 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, 0);
1483 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, 0);
1484 break;
1485 case 2:
1486 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_2x[0]);
1487 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_2x[1]);
1488 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_2x[2]);
1489 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_2x[3]);
1490 break;
1491 case 4:
1492 radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_4x[0]);
1493 radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_4x[1]);
1494 radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_4x[2]);
1495 radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_4x[3]);
1496 break;
1497 case 8:
1498 radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
1499 radeon_emit(cs, cm_sample_locs_8x[0]);
1500 radeon_emit(cs, cm_sample_locs_8x[4]);
1501 radeon_emit(cs, 0);
1502 radeon_emit(cs, 0);
1503 radeon_emit(cs, cm_sample_locs_8x[1]);
1504 radeon_emit(cs, cm_sample_locs_8x[5]);
1505 radeon_emit(cs, 0);
1506 radeon_emit(cs, 0);
1507 radeon_emit(cs, cm_sample_locs_8x[2]);
1508 radeon_emit(cs, cm_sample_locs_8x[6]);
1509 radeon_emit(cs, 0);
1510 radeon_emit(cs, 0);
1511 radeon_emit(cs, cm_sample_locs_8x[3]);
1512 radeon_emit(cs, cm_sample_locs_8x[7]);
1513 break;
1514 case 16:
1515 radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 16);
1516 radeon_emit(cs, cm_sample_locs_16x[0]);
1517 radeon_emit(cs, cm_sample_locs_16x[4]);
1518 radeon_emit(cs, cm_sample_locs_16x[8]);
1519 radeon_emit(cs, cm_sample_locs_16x[12]);
1520 radeon_emit(cs, cm_sample_locs_16x[1]);
1521 radeon_emit(cs, cm_sample_locs_16x[5]);
1522 radeon_emit(cs, cm_sample_locs_16x[9]);
1523 radeon_emit(cs, cm_sample_locs_16x[13]);
1524 radeon_emit(cs, cm_sample_locs_16x[2]);
1525 radeon_emit(cs, cm_sample_locs_16x[6]);
1526 radeon_emit(cs, cm_sample_locs_16x[10]);
1527 radeon_emit(cs, cm_sample_locs_16x[14]);
1528 radeon_emit(cs, cm_sample_locs_16x[3]);
1529 radeon_emit(cs, cm_sample_locs_16x[7]);
1530 radeon_emit(cs, cm_sample_locs_16x[11]);
1531 radeon_emit(cs, cm_sample_locs_16x[15]);
1532 break;
1533 }
1534 }
1535
1536 static void radv_cayman_get_sample_position(struct radv_device *device,
1537 unsigned sample_count,
1538 unsigned sample_index, float *out_value)
1539 {
1540 int offset, index;
1541 struct {
1542 int idx:4;
1543 } val;
1544 switch (sample_count) {
1545 case 1:
1546 default:
1547 out_value[0] = out_value[1] = 0.5;
1548 break;
1549 case 2:
1550 offset = 4 * (sample_index * 2);
1551 val.idx = (eg_sample_locs_2x[0] >> offset) & 0xf;
1552 out_value[0] = (float)(val.idx + 8) / 16.0f;
1553 val.idx = (eg_sample_locs_2x[0] >> (offset + 4)) & 0xf;
1554 out_value[1] = (float)(val.idx + 8) / 16.0f;
1555 break;
1556 case 4:
1557 offset = 4 * (sample_index * 2);
1558 val.idx = (eg_sample_locs_4x[0] >> offset) & 0xf;
1559 out_value[0] = (float)(val.idx + 8) / 16.0f;
1560 val.idx = (eg_sample_locs_4x[0] >> (offset + 4)) & 0xf;
1561 out_value[1] = (float)(val.idx + 8) / 16.0f;
1562 break;
1563 case 8:
1564 offset = 4 * (sample_index % 4 * 2);
1565 index = (sample_index / 4) * 4;
1566 val.idx = (cm_sample_locs_8x[index] >> offset) & 0xf;
1567 out_value[0] = (float)(val.idx + 8) / 16.0f;
1568 val.idx = (cm_sample_locs_8x[index] >> (offset + 4)) & 0xf;
1569 out_value[1] = (float)(val.idx + 8) / 16.0f;
1570 break;
1571 case 16:
1572 offset = 4 * (sample_index % 4 * 2);
1573 index = (sample_index / 4) * 4;
1574 val.idx = (cm_sample_locs_16x[index] >> offset) & 0xf;
1575 out_value[0] = (float)(val.idx + 8) / 16.0f;
1576 val.idx = (cm_sample_locs_16x[index] >> (offset + 4)) & 0xf;
1577 out_value[1] = (float)(val.idx + 8) / 16.0f;
1578 break;
1579 }
1580 }
1581
1582 void radv_device_init_msaa(struct radv_device *device)
1583 {
1584 int i;
1585 radv_cayman_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
1586
1587 for (i = 0; i < 2; i++)
1588 radv_cayman_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
1589 for (i = 0; i < 4; i++)
1590 radv_cayman_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
1591 for (i = 0; i < 8; i++)
1592 radv_cayman_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
1593 for (i = 0; i < 16; i++)
1594 radv_cayman_get_sample_position(device, 16, i, device->sample_locations_16x[i]);
1595 }